A Re-Parameterized Lightweight Residual Attention Framework for Resource-Constrained Edge Computing

Edge vision systems require convolutional neural networks (CNNs) that preserve recognition accuracy under strict storage, computation, and latency constraints. Although ResNet18 is a compact residual backbone, direct deployment on resource-constrained devices remains costly, whereas simple channel reduction weakens representation capacity. This study aims to build a deployable ResNet18-based classifier that reduces model complexity while recovering the accuracy lost during compression. We propose a lightweight framework that combines global channel scaling, a re-parameterized attention residual block, and teacher–student knowledge distillation. The proposed block uses multi-branch convolution and squeeze-and-excitation attention during training, then folds the linear branches into a single 3-by-3 convolution for inference. Experiments on CIFAR-100 show that the final model reduces parameters from 11.220 M to 2.841 M, retains comparable Top-1 accuracy (0.7579 vs. 0.7606), improves Top-5 accuracy (0.9340 vs. 0.9253), and reduces graphics processing unit (GPU) batch inference latency from 3.279 ms to 2.161 ms. Deployment on PYNQ-Z2 verifies the complete camera-based CPU-side inference workflow, with an average end-to-end latency of 421.467 ms/frame. The results indicate that residual topology preservation, re-parameterized feature enhancement, and distillation form a practical route for edge-oriented lightweight CNN deployment.